claude-skill（Claude 技能）

Claude Code Headless Mode

Use this skill when the job should be executed through Claude Code itself, not solved inline. Focus on commands and workflows that match current stable Claude Code behavior.

Core Rules

• Treat claude --help on the target machine as the compatibility floor. CLI flags move faster than blog posts and copied examples.
• Permission rule syntax varies by build. Anthropic docs often show forms like Bash(git diff *), while the installed claude 2.1.71 help on this machine still shows Bash(git:*). Mirror the syntax shown by the target machine's claude --help.
• Default to the model the user already configured through /model, settings, or ANTHROPIC_MODEL.
• Do not add --model unless the user explicitly asked for a model override or the workflow must pin a model for reproducibility.
• Prefer --append-system-prompt over --system-prompt unless replacing the default Claude Code behavior is intentional.
• Default to safe automation. If the user wants a truly unattended run, use explicit permission rules or dontAsk; reserve bypassPermissions for isolated environments.

Quick Verification

Run these checks before giving advanced advice:

claude --version
claude auth status --text
claude --help

If installation or updates look odd, use:

claude doctor

Installation Guidance

• Anthropic's getting-started docs still show npm install -g @anthropic-ai/claude-code as the standard install path.
• Newer builds may also support native installer flows and claude install.
• If the user needs installation help, point them to the official Claude Code setup docs and verify the result with claude doctor rather than assuming one installer path is universal.

Headless Command Patterns

Basic Non-Interactive Run

Use -p / --print for non-interactive execution:

claude -p "summarize the repository architecture"

Add --output-format json when the caller needs machine-readable output:

claude -p "review the auth layer for risks" --output-format json

Model Selection

• Omit --model by default.
• If the user explicitly wants a model override, use claude --model <alias-or-name>....
• For persistent defaults, prefer settings (model) or ANTHROPIC_MODEL.
• For third-party deployments, pin models via settings or environment variables instead of bolting --model onto every command example.

Permission Modes

Claude Code supports these permission modes:

Important clarifications:

• acceptEdits is the skill's recommended default, not the CLI default.
• If the user says "no prompts at all," prefer permission rules or dontAsk with explicit allow rules.
• Only recommend bypassPermissions when the environment is already isolated and the user accepts the risk.
• For read-only analysis, prefer --tools plus default or plan; do not reach for bypassPermissions just to suppress prompts.

Tool Availability vs Permission Approval

Do not mix these up:

• --tools restricts which built-in tools are available at all.
• --allowedTools pre-approves specific tools or tool rules so Claude does not prompt for them.
• --disallowedTools removes tools or rules from context.

Permission rules follow Tool or Tool(specifier) syntax.

Use wildcard rules when the command will include arguments:

• Good: Bash(git diff *)
• Good: Bash(npm run test *)
• Risky for real use: Bash(find) because it matches only the exact literal command find

If the local CLI help shows colon syntax such as Bash(find:*), use that form on that machine. The important part is to allow an argument-aware rule rather than an exact literal command.

If the user wants Claude limited to a narrow tool family, you should usually use both --tools and --allowedTools: --tools defines the hard boundary, --allowedTools removes prompts inside that boundary.

Output Formats

• text: default human-readable output
• json: one final structured result
• stream-json: event stream for long-running automation

Do not promise a fixed JSON schema unless you have validated it on the target version. Prefer wording like "returns a final result object with response text, timing, and session metadata."

Commonly Safe Flags

These are appropriate starting points on current stable builds:

• --append-system-prompt
• --allowedTools
• --disallowedTools
• --tools
• --permission-mode
• --output-format
• --mcp-config
• --continue / --resume
• --settings / --setting-sources
• --session-id
• --add-dir
• --max-budget-usd
• --fallback-model for print mode

Version-Sensitive Flags

Published docs sometimes mention flags that are absent from the installed binary on a given machine. Before emitting less-common flags, verify them with claude --help.

Recommended Command Templates

Read-Only Analysis

claude -p "count the total lines of code in this repo, grouped by language" \
  --permission-mode default \
  --tools "Bash,Read" \
  --allowedTools "Read" "Bash(find:*)" "Bash(wc:*)"

Safe Edit Run

claude -p "fix the failing login test and rerun the relevant test command" \
  --permission-mode acceptEdits \
  --tools "Bash,Read,Edit,Write" \
  --allowedTools "Read" "Edit" "Write" "Bash(npm test *)"

JSON Report

claude -p "review the repository for security issues and produce a concise report" \
  --output-format json \
  --append-system-prompt "Focus on concrete findings, exploitability, and mitigations."

Resume a Session

claude -r "$session_id" -p "continue by updating the tests and summarizing what changed"

Continue the Most Recent Session

claude -c -p "continue with the next step"

Use MCP Tools

claude -p "investigate the API latency spike" \
  --permission-mode acceptEdits \
  --mcp-config monitoring-tools.json \
  --allowedTools "Read" "mcp__datadog__*" "mcp__prometheus__*"

Fully Unattended Execution in an Isolated Environment

Use only when the environment is already sandboxed:

claude -p "run the migration and fix the resulting type errors" \
  --permission-mode bypassPermissions \
  --tools "Bash,Read,Edit,Write"

Execution Workflow

1. Confirm the user wants Claude Code CLI rather than an inline answer.
2. Verify the installed CLI shape with claude --help if you plan to use uncommon flags.
3. Choose the least-permissive mode that still fits the task.
4. Build the command without --model unless the user explicitly asked for an override.
5. Restrict tools with --tools when safety or predictability matters.
6. Use --allowedTools for prompt-free approval of known-safe actions.
7. Return exact commands plus short caveats about assumptions and safety.

When To Pause

Pause only when one of these is materially unclear:

• The user wants a specific model or provider behavior that requires pinning.
• They asked for a fully unattended run in an environment that is not clearly sandboxed.
• The workflow depends on a Claude Code feature that is not visible in claude --help.

Otherwise, proceed and give the best current command.

Final Response Expectations

When using this skill, the output should usually include:

• the exact command or command sequence
• a one-line note that the configured Claude model is being used by default
• any permission or isolation caveat
• the resume command if the workflow is meant to continue later

References

• references/examples.md for extended patterns